[llvm] [AMDGPU] Introduce "amdgpu-uniform-intrinsic-combine" pass to combine uniform AMDGPU lane Intrinsics. (PR #116953)

Sameer Sahasrabuddhe via llvm-commits llvm-commits at lists.llvm.org
Thu Mar 6 20:27:45 PST 2025


================
@@ -0,0 +1,535 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -amdgpu-enable-uniform-intrinsic-combine=0 -O3 -S < %s | FileCheck %s -check-prefix=CURRENT-CHECK
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -passes=amdgpu-uniform-intrinsic-combine -S < %s | FileCheck %s -check-prefix=PASS-CHECK
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -passes=amdgpu-uniform-intrinsic-combine,early-cse,instcombine,simplifycfg -S < %s | FileCheck %s -check-prefix=DCE-CHECK
+
+define protected amdgpu_kernel void @trivial_waterfall_eq_zero(ptr addrspace(1) %out) {
+; CURRENT-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero(
+; CURRENT-CHECK-SAME: ptr addrspace(1) writeonly captures(none) [[OUT:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+; CURRENT-CHECK-NEXT:  [[ENTRY:.*:]]
+; CURRENT-CHECK-NEXT:    [[TMP0:%.*]] = tail call i32 @llvm.amdgcn.ballot.i32(i1 true)
+; CURRENT-CHECK-NEXT:    [[IS_DONE_PEEL:%.*]] = icmp eq i32 [[TMP0]], 0
+; CURRENT-CHECK-NEXT:    br i1 [[IS_DONE_PEEL]], label %[[EXIT:.*]], label %[[IF_PEEL:.*]]
+; CURRENT-CHECK:       [[IF_PEEL]]:
+; CURRENT-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; CURRENT-CHECK-NEXT:    br label %[[EXIT]]
+; CURRENT-CHECK:       [[EXIT]]:
+; CURRENT-CHECK-NEXT:    ret void
+;
+; PASS-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero(
+; PASS-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0:[0-9]+]] {
+; PASS-CHECK-NEXT:  [[ENTRY:.*]]:
+; PASS-CHECK-NEXT:    br label %[[WHILE:.*]]
+; PASS-CHECK:       [[WHILE]]:
+; PASS-CHECK-NEXT:    [[DONE:%.*]] = phi i1 [ false, %[[ENTRY]] ], [ true, %[[IF:.*]] ]
+; PASS-CHECK-NEXT:    [[NOT_DONE:%.*]] = xor i1 [[DONE]], true
+; PASS-CHECK-NEXT:    [[BALLOT:%.*]] = tail call i64 @llvm.amdgcn.ballot.i64(i1 [[NOT_DONE]])
+; PASS-CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[NOT_DONE]], true
+; PASS-CHECK-NEXT:    [[IS_DONE:%.*]] = icmp eq i64 [[BALLOT]], 0
+; PASS-CHECK-NEXT:    br i1 [[TMP0]], label %[[EXIT:.*]], label %[[IF]]
+; PASS-CHECK:       [[IF]]:
+; PASS-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; PASS-CHECK-NEXT:    br label %[[WHILE]]
+; PASS-CHECK:       [[EXIT]]:
+; PASS-CHECK-NEXT:    ret void
+;
+; DCE-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero(
+; DCE-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0:[0-9]+]] {
+; DCE-CHECK-NEXT:  [[ENTRY:.*:]]
+; DCE-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; DCE-CHECK-NEXT:    ret void
+;
+entry:
+  br label %while
+
+while:
+  %done = phi i1 [ 0, %entry ], [ 1, %if ]
+  %not_done = xor i1 %done, true
+  %ballot = tail call i64 @llvm.amdgcn.ballot.i64(i1 %not_done)
+  %is_done = icmp eq i64 %ballot, 0 ; in this case is_done = !not_done
+  br i1 %is_done, label %exit, label %if
+
+if:
+  store i32 5, ptr addrspace(1) %out
+  br label %while
+
+exit:
+  ret void
+}
+
+define protected amdgpu_kernel void @trivial_waterfall_eq_zero_swap_op(ptr addrspace(1) %out) {
+; CURRENT-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero_swap_op(
+; CURRENT-CHECK-SAME: ptr addrspace(1) writeonly captures(none) [[OUT:%.*]]) local_unnamed_addr #[[ATTR0]] {
+; CURRENT-CHECK-NEXT:  [[ENTRY:.*:]]
+; CURRENT-CHECK-NEXT:    [[TMP0:%.*]] = tail call i32 @llvm.amdgcn.ballot.i32(i1 true)
+; CURRENT-CHECK-NEXT:    [[IS_DONE_PEEL:%.*]] = icmp eq i32 [[TMP0]], 0
+; CURRENT-CHECK-NEXT:    br i1 [[IS_DONE_PEEL]], label %[[EXIT:.*]], label %[[IF_PEEL:.*]]
+; CURRENT-CHECK:       [[IF_PEEL]]:
+; CURRENT-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; CURRENT-CHECK-NEXT:    br label %[[EXIT]]
+; CURRENT-CHECK:       [[EXIT]]:
+; CURRENT-CHECK-NEXT:    ret void
+;
+; PASS-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero_swap_op(
+; PASS-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0]] {
+; PASS-CHECK-NEXT:  [[ENTRY:.*]]:
+; PASS-CHECK-NEXT:    br label %[[WHILE:.*]]
+; PASS-CHECK:       [[WHILE]]:
+; PASS-CHECK-NEXT:    [[DONE:%.*]] = phi i1 [ false, %[[ENTRY]] ], [ true, %[[IF:.*]] ]
+; PASS-CHECK-NEXT:    [[NOT_DONE:%.*]] = xor i1 [[DONE]], true
+; PASS-CHECK-NEXT:    [[BALLOT:%.*]] = tail call i64 @llvm.amdgcn.ballot.i64(i1 [[NOT_DONE]])
+; PASS-CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[NOT_DONE]], true
+; PASS-CHECK-NEXT:    [[IS_DONE:%.*]] = icmp eq i64 0, [[BALLOT]]
+; PASS-CHECK-NEXT:    br i1 [[TMP0]], label %[[EXIT:.*]], label %[[IF]]
+; PASS-CHECK:       [[IF]]:
+; PASS-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; PASS-CHECK-NEXT:    br label %[[WHILE]]
+; PASS-CHECK:       [[EXIT]]:
+; PASS-CHECK-NEXT:    ret void
+;
+; DCE-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_eq_zero_swap_op(
+; DCE-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0]] {
+; DCE-CHECK-NEXT:  [[ENTRY:.*:]]
+; DCE-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; DCE-CHECK-NEXT:    ret void
+;
+entry:
+  br label %while
+
+while:
+  %done = phi i1 [ 0, %entry ], [ 1, %if ]
+  %not_done = xor i1 %done, true
+  %ballot = tail call i64 @llvm.amdgcn.ballot.i64(i1 %not_done)
+  %is_done = icmp eq i64 0, %ballot ; in this case is_done = !not_done
+  br i1 %is_done, label %exit, label %if
+
+if:
+  store i32 5, ptr addrspace(1) %out
+  br label %while
+
+exit:
+  ret void
+}
+
+define protected amdgpu_kernel void @trivial_waterfall_ne_one(ptr addrspace(1) %out) {
+; CURRENT-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_ne_one(
+; CURRENT-CHECK-SAME: ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1:[0-9]+]] {
+; CURRENT-CHECK-NEXT:  [[ENTRY:.*:]]
+; CURRENT-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; CURRENT-CHECK-NEXT:    br label %[[WHILE:.*]]
+; CURRENT-CHECK:       [[WHILE]]:
+; CURRENT-CHECK-NEXT:    [[TMP0:%.*]] = tail call i32 @llvm.amdgcn.ballot.i32(i1 true)
+; CURRENT-CHECK-NEXT:    [[IS_DONE_NOT:%.*]] = icmp eq i32 [[TMP0]], 1
+; CURRENT-CHECK-NEXT:    br i1 [[IS_DONE_NOT]], label %[[EXIT:.*]], label %[[WHILE]], !llvm.loop [[LOOP0:![0-9]+]]
+; CURRENT-CHECK:       [[EXIT]]:
+; CURRENT-CHECK-NEXT:    ret void
+;
+; PASS-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_ne_one(
+; PASS-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0]] {
+; PASS-CHECK-NEXT:  [[ENTRY:.*]]:
+; PASS-CHECK-NEXT:    br label %[[WHILE:.*]]
+; PASS-CHECK:       [[WHILE]]:
+; PASS-CHECK-NEXT:    [[DONE:%.*]] = phi i1 [ false, %[[ENTRY]] ], [ true, %[[IF:.*]] ]
+; PASS-CHECK-NEXT:    [[BALLOT:%.*]] = tail call i64 @llvm.amdgcn.ballot.i64(i1 [[DONE]])
+; PASS-CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[DONE]], true
+; PASS-CHECK-NEXT:    [[IS_DONE:%.*]] = icmp ne i64 [[BALLOT]], 1
+; PASS-CHECK-NEXT:    br i1 [[TMP0]], label %[[IF]], label %[[EXIT:.*]]
+; PASS-CHECK:       [[IF]]:
+; PASS-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; PASS-CHECK-NEXT:    br label %[[WHILE]]
+; PASS-CHECK:       [[EXIT]]:
+; PASS-CHECK-NEXT:    ret void
+;
+; DCE-CHECK-LABEL: define protected amdgpu_kernel void @trivial_waterfall_ne_one(
+; DCE-CHECK-SAME: ptr addrspace(1) [[OUT:%.*]]) #[[ATTR0]] {
+; DCE-CHECK-NEXT:  [[ENTRY:.*:]]
+; DCE-CHECK-NEXT:    store i32 5, ptr addrspace(1) [[OUT]], align 4
+; DCE-CHECK-NEXT:    ret void
+;
+entry:
+  br label %while
+
+while:
+  %done = phi i1 [ 0, %entry ], [ 1, %if ]
+  %ballot = tail call i64 @llvm.amdgcn.ballot.i64(i1 %done)
+  %is_done = icmp ne i64 %ballot, 1 ; in this case is_done = !done
----------------
ssahasra wrote:

I am not sure how you reach that conclusion. To ask the same question again, what does "ballot is not 1" mean?

https://github.com/llvm/llvm-project/pull/116953


More information about the llvm-commits mailing list